Tests show OpenAI's textGrain watermarks weaken after editing
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Marking SI text is possible, but short or edited passages cannot be reliably checked with a watermark alone; transparency requires several independent methods.
TextGrain changes the statistics of token selection using a secret key and treats entropy as a budget, allowing developers to adjust the trade-off between watermark detectability and text variety. In OpenAI's own evaluations, the detector found the watermark in about 80% of 200-token passages and 95% of 400-token passages. Replacing 10% of words with synonyms in 400-token text reduced detection from 92% to 66%; replacing 25% reduced it to 17%. The estimated false-positive rate is 1%. Article 50 of the European AI Act requires machine-readable marking of synthetic text but sets no numerical robustness threshold. The voluntary code of practice does not require watermarks on passages shorter than 200 tokens. Researchers interviewed by IEEE consider the early methodology promising, but independent researchers are only now gaining access for evaluation. OpenAI emphasises that watermarks should complement other content-provenance methods.
Sources 1